Yes! The way those physics models are created is so cool. The article somewhat explains it, but it’s mostly a fluff-piece for things unrelated to genAI. More in-depth:
The physically accurate simulation is great but slow. So we can create a neural network (there’s a huge variety in shapes), and give it an example of physics, and tell it to make a guess as to what it’ll look like in, say, 1ms. We make it improve at this billions of times, and eventually it becomes “good enough” in most cases. By doing those 1ms steps in a loop, we get a full simulation. Because we chose the shape of it, we can pick a shape that’s quite fast to compute, and now we have a less-accurate but faster simulation.
The really cool thing is that sometimes, these models are better than the more expensive physics simulation, probably because real physics is logical and logical things are easier to learn.
We’ve done things like this for ages. One way we can improve them is by giving them multiple time steps. Unfortunately they kinda suck at seeing connections over time, so this is expensive. Luckily, transformers were invented! This is a neural network shape that is really good at seeing connections over one dimension, like time, while still being pretty cheap and really easy to do run in parallel (which is how you can go fast nowadays).
With a bunch of extra wiring, transformers also become GPT, i.e. text-based AIs. That’s why they suddenly got way better; they went from being able to see connections with words maybe 3-4 steps back, to recently a literal million. This is basically the only relationship with “AI” this has.
Yes! The way those physics models are created is so cool. The article somewhat explains it, but it’s mostly a fluff-piece for things unrelated to genAI. More in-depth:
The physically accurate simulation is great but slow. So we can create a neural network (there’s a huge variety in shapes), and give it an example of physics, and tell it to make a guess as to what it’ll look like in, say, 1ms. We make it improve at this billions of times, and eventually it becomes “good enough” in most cases. By doing those 1ms steps in a loop, we get a full simulation. Because we chose the shape of it, we can pick a shape that’s quite fast to compute, and now we have a less-accurate but faster simulation.
The really cool thing is that sometimes, these models are better than the more expensive physics simulation, probably because real physics is logical and logical things are easier to learn.
We’ve done things like this for ages. One way we can improve them is by giving them multiple time steps. Unfortunately they kinda suck at seeing connections over time, so this is expensive. Luckily, transformers were invented! This is a neural network shape that is really good at seeing connections over one dimension, like time, while still being pretty cheap and really easy to do run in parallel (which is how you can go fast nowadays).
With a bunch of extra wiring, transformers also become GPT, i.e. text-based AIs. That’s why they suddenly got way better; they went from being able to see connections with words maybe 3-4 steps back, to recently a literal million. This is basically the only relationship with “AI” this has.